Three Boxes, and None of Them Are the Question

TL;DR: Gabriela Pereira polled the DIY MFA audience with three options: AI User, AI-Cautious, AI-Avoidant. She puts herself in the middle box and names her reasons, unlicensed training data and the environmental cost, both of which are real. Her practical line is the best sentence in the piece: putting your fingers in your ears about AI is “about as practical as being anti-internet in the 1990s.” I would still not check any of the three boxes, because all three are answers about stance and the only thing that reaches the page is authorship.

The question is on the wrong axis.

Not because it is a bad question. Because it measures how you feel, and feelings do not survive the trip to the page. Nothing about your position on a technology is audible in a paragraph. What is audible is whether somebody chose the words.

What she is actually asking

Pereira’s poll sorts writers three ways. AI User: uses it regularly to save time on tasks they dislike. AI-Cautious: wants more information before committing. AI-Avoidant: opposed, spellcheck excepted.

She takes the middle box herself, and she is specific about why: “I don’t like how many LLMs use unlicensed work to train their models, nor do I like what these large computer farms are doing to our environment.”

Both objections are real and I am not going to pretend they are not. The training data question is a genuine grievance with real people on the wrong end of it, and the Anthropic settlement, which she calls “not great, but it’s better than nothing,” is a fair reading of a fair complaint. The energy question is real. Anyone who tells you those concerns are manufactured is selling something.

Her advice is sound too and I would repeat all of it. Read the terms of any tool before you feed it anything. Do not put a full manuscript into a service without knowing what ownership you keep. Verify every output independently, because garbage in stays garbage out and nothing in the loop is checking. Consult the Authors Guild materials. That is a good list and most writers have done none of it.

Where the categories break

Take three writers.

One is AI-Avoidant, on principle, loudly. She writes to a template she learned in a workshop in 2011 and has not questioned since. Every chapter opens with weather. Every emotional beat is named on arrival. Nothing in it was decided, it was all defaulted, and defaults are just somebody else’s choices you stopped noticing.

One is an AI User. He drafts with the machine, then rewrites every paragraph until the rhythm is his, cuts the tidy endings it keeps reaching for, and puts back the strange thing it kept smoothing out. What ships is his.

One is AI-Cautious. She has not decided. In the meantime she is writing, carefully, choosing every sentence.

Two of those three are making authored work. The poll puts one of those two in the same box as the person who is not.

That is the failure. The categories sort by input and the thing that matters is output, and there is no reliable relationship between them.

The only question

You can hear it. That is not a metaphor and I have said it in this space before, so I will not relitigate the whole argument here. There is a texture to prose where somebody was choosing, and a different texture where the choosing stopped and something plausible got accepted.

The plausible sentence is the hazard. It is grammatical, it is on-topic, it arrives at the right length, and it is nobody’s. It has no fingerprint because no hand closed on it. That sentence can come out of a model at 4am. It can also come out of a tired writer reaching for the phrasing that always works, and it sounds identical in both cases, which should tell you something about where the problem actually lives.

So the question I would put on the poll, replacing all three boxes:

Did you choose this, or did you accept it?

Line by line. That is the whole gate. It is the same gate it was before any of this existed. What changed is that accepting is now frictionless, instant, and available in unlimited quantity, so a habit that used to cost you an afternoon of laziness now costs you a book.

Where this house stands, since you asked

We use it. Openly, with names on the work, and we are not going to be coy about that in a piece about how other people should feel.

I think the affirmative case is stronger than the defensive one and I would rather make it than hedge. The tools are genuinely good at the parts of this work that were never the art: the third pass on a structure, the research spadework, the format conversion, the ninety minutes of formatting that used to eat an evening. Getting that time back and spending it on the sentence is not a compromise, it is a better allocation of a writer’s life. Pereira’s own framing supports this more than her box does, because the 1990s comparison is exactly right and nobody now thinks the anti-internet writers were protecting their craft.

And the objections stay live at the same time. Both things. The training data question deserves a settlement better than the one it got. The energy question deserves a real answer rather than a press release. Holding those two positions at once is not fence-sitting, it is just the accurate description of a technology that is both useful and unfinished.

What I will not do is grade a manuscript on which box its author ticked. I grade it on whether anyone was home while it was written.

The poll asks how you feel. The page asks who decided.

Only one of those leaves a mark.

Frequently asked questions

Is there a way to tell from prose alone whether a tool was involved?

Not reliably, and the people most confident that they can are wrong often enough to make it dangerous as a standard. What the supposed tells actually detect is unedited output, which is a signal about how much care went in rather than about which software was open. A heavily revised draft will not trip them and a tired writer’s unrevised prose frequently will, which is why the useful question is about authorship rather than tooling.

Are the training-data and energy objections reasons not to use these tools?

They are reasons to be honest about the cost of using them, which is a different obligation. Pereira holds the objections and uses the tools cautiously anyway, and that is a coherent position rather than a contradiction. The incoherent version is the one where somebody uses the tools daily, benefits from them, and performs opposition in public because the performance is cheaper than the disclosure.

What should a writer actually do before putting work into an AI tool?

Read the terms and find out what happens to what you paste, specifically whether it can be retained or used for training and whether you keep ownership of the output. Do not put an unpublished full manuscript into a service you have not checked on that point. Verify any factual claim it hands you against a source you found yourself. That is the same diligence you would apply to any vendor holding your work, and almost nobody performs it because the interface does not feel like a vendor.